Modeling ALS Progression¶

In [129]:
import pandas as pd
import numpy as np
from keras.layers import (
    MultiHeadAttention,
    Dense,
    Dropout,
    LayerNormalization,
    Input,
    GlobalAveragePooling1D,
)
from keras.models import Model
from keras.optimizers.legacy import RMSprop
from sklearn.model_selection import KFold
from keras.callbacks import Callback, ModelCheckpoint
import os
from sklearn.metrics import (
    mean_squared_error,
    root_mean_squared_error,
    mean_absolute_error,
    r2_score,
)
import json
import warnings
import matplotlib.pyplot as plt
import seaborn as sbn

plt.rcParams["font.family"] = "IBM Plex Mono"

warnings.filterwarnings("ignore", category=Warning)
In [130]:
df = pd.read_csv('Data/preprocessed/data.csv')
df
Out[130]:
subject_id feature 0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 ... 111.0 112.0 113.0 114.0 115.0 116.0 117.0 118.0 119.0 120.0
0 1492 Q1_Speech 4.0 4.0 4.0 4.0 4.0 4.0 3.0 4.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
1 1524 Q1_Speech 4.0 3.0 4.0 4.0 4.0 4.0 3.0 4.0 ... 4.0 4.0 4.0 4.0 4.0 4.0 4.0 4.0 4.0 4.0
2 2018 Q1_Speech 4.0 4.0 4.0 4.0 4.0 4.0 4.0 4.0 ... 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0
3 2025 Q1_Speech 4.0 4.0 4.0 4.0 4.0 4.0 4.0 4.0 ... 4.0 4.0 4.0 4.0 4.0 4.0 4.0 4.0 4.0 4.0
4 2194 Q1_Speech 4.0 4.0 4.0 4.0 4.0 4.0 4.0 4.0 ... 3.0 3.0 3.0 3.0 3.0 3.0 3.0 3.0 3.0 3.0
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
38935 996984 Riluzole 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
38936 997561 Riluzole 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
38937 997649 Riluzole 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
38938 998278 Riluzole 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 ... 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0
38939 999344 Riluzole 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

38940 rows × 123 columns

In [131]:
def dataframe_to_data(df):
    data = []
    for sid in df.subject_id.unique():
        data.append(df[df.subject_id == sid].values[:, 2:])
    return np.array(data)
In [132]:
data = dataframe_to_data(df)
print(f'Data Shape :: {data.shape}')

Data Shape :: (1770, 22, 121)
In [133]:
def downsample_to_monthly(data, steps_per_month=7):
    monthly_data = []
    for i in range(15):
        start = i * steps_per_month
        end = start + steps_per_month
        month_avg = np.mean(data[:, :, start:end], axis=2)
        monthly_data.append(month_avg)
    month16_avg = np.mean(data[:, :, 105:], axis=2)
    monthly_data.append(month16_avg)
    monthly_data = np.stack(monthly_data, axis=2)
    return np.transpose(monthly_data.astype(np.float32), (0, 2, 1))
In [134]:
data = downsample_to_monthly(data)
print(f'Monthly Data Shape :: {data.shape}')

Monthly Data Shape :: (1770, 16, 22)
In [135]:
def generate_patient_rolling_data(
    data, static_feature_indices=[18, 19, 20, 21], input_steps=4, output_offset=1
):
    num_features = data.shape[1]
    total_steps = data.shape[0]
    all_indices = np.arange(num_features)
    temporal_indices = np.setdiff1d(all_indices, static_feature_indices)
    max_start = total_steps - input_steps - output_offset + 1
    X_list, Y_list = [], []
    for t in range(max_start):
        x_window = data[t : t + input_steps, :]
        y_target = data[t + input_steps + output_offset - 1, temporal_indices]
        static_values = data[0, static_feature_indices]
        static_repeated = np.repeat(static_values[np.newaxis, :], input_steps, axis=0)
        x_combined = np.concatenate(
            [x_window[:, temporal_indices], static_repeated], axis=1
        )
        X_list.append(x_combined)
        Y_list.append(y_target)
    return X_list, Y_list
In [136]:
def generate_multistep_data(data):
    X_list, Y_list = [], []
    for p in range(data.shape[0]):
        X_combined, y_target = generate_patient_rolling_data(data[p])
        X_list.extend(X_combined)
        Y_list.extend(y_target)
    return np.array(X_list), np.array(Y_list)
In [137]:
def transformer_encoder(inputs, head_size=64, num_heads=4, ff_dim=128, dropout=0.1):
    x_ = MultiHeadAttention(key_dim=head_size, num_heads=num_heads, dropout=dropout)(
        inputs, inputs
    )
    x_ = Dropout(dropout)(x_)
    x_ = LayerNormalization(epsilon=1e-6)(x_)
    res = x_ + inputs
    x_ = Dense(ff_dim, activation="relu")(res)
    x_ = Dropout(dropout)(x_)
    x_ = Dense(inputs.shape[-1])(x_)
    x_ = LayerNormalization(epsilon=1e-6)(x_)
    return x_ + res
In [138]:
def build_transformer_model():
    inputs = Input(shape=(4, 22))
    x_ = transformer_encoder(inputs, head_size=64, num_heads=4, ff_dim=128, dropout=0.1)
    x_ = GlobalAveragePooling1D()(x_)
    x_ = Dense(128, activation="relu")(x_)
    x_ = Dropout(0.3)(x_)
    outputs = Dense(18)(x_)
    model = Model(inputs, outputs)
    model.compile(optimizer=RMSprop(learning_rate=0.001), loss='mse')
    return model
In [139]:
def autoregressive_forecast(
    data,
    model,
    input_steps=4,
    static_feature_indices=[18, 19, 20, 21],
):
    num_features = data.shape[2]
    temporal_indices = np.setdiff1d(np.arange(num_features), static_feature_indices)
    static_feats = data[:, 0, static_feature_indices]
    x_seq = data[:, :input_steps, :]
    y_preds_all = []
    for _ in range(input_steps):
        static_repeated = np.repeat(static_feats[:, np.newaxis, :], input_steps, axis=1)
        x_input = np.concatenate(
            [x_seq[:, :, temporal_indices], static_repeated], axis=2
        )
        y_pred = model.predict(x_input, verbose=0)
        y_preds_all.append(y_pred)
        next_step = np.concatenate([y_pred, static_feats], axis=1)
        x_seq = np.concatenate([x_seq[:, 1:, :], next_step[:, np.newaxis, :]], axis=1)
    return np.stack(y_preds_all, axis=1)
In [140]:
def evaluate_by_horizon(data, model, horizons=[3, 6, 9, 12]):
    y_preds = autoregressive_forecast(data, model)
    y_true = np.stack([data[:, 3 + h, :18] for h in range(4)], axis=1)
    mse_per_horizon = {}
    for i, h in enumerate(horizons):
        mse = mean_squared_error(y_true[:, i], y_preds[:, i])
        mse_per_horizon[f"{h}m"] = mse
    return mse_per_horizon
In [141]:
class TrainingCallback(Callback):
    def __init__(self, train_idx, test_idx):
        self.train_idx = train_idx
        self.test_idx = test_idx
        self.history = {
            "train": {"3m": [], "6m": [], "9m": [], "12m": []},
            "test": {"3m": [], "6m": [], "9m": [], "12m": []},
        }

    def on_epoch_end(self, epoch, logs=None):
        print(f"🟩 Epoch {epoch + 1} Evaluation:")
        train_data = data[self.train_idx]
        test_data = data[self.test_idx]

        train_mse = evaluate_by_horizon(train_data, self.model)
        test_mse = evaluate_by_horizon(test_data, self.model)
        for h in ["3m", "6m", "9m", "12m"]:
            self.history["train"][h].append(train_mse[h])
            self.history["test"][h].append(test_mse[h])
        print(
            "  🔹 Train MSEs: "
            + ", ".join([f"{k}: {v:.5f}" for k, v in train_mse.items()])
        )
        print(
            "  🔸 Test  MSEs: "
            + ", ".join([f"{k}: {v:.5f}" for k, v in test_mse.items()])
        )
In [142]:
kfold = KFold(n_splits=5, shuffle=True, random_state=1)
history = {}
for i, (train_idx, test_idx) in enumerate(kfold.split(np.zeros((data.shape[0], 1)))):
    print(f"[INFO] Working on Fold {i+1}")
    model = build_transformer_model()
    model_dir = f"models/fold_{i+1}"
    os.makedirs(model_dir, exist_ok=True)
    callbacks = [
        ModelCheckpoint(
            f"{model_dir}/model.h5",
            save_best_only=True,
            save_weights_only=True,
            verbose=False,
        ),
        TrainingCallback(train_idx, test_idx),
    ]
    train_x, train_y = generate_multistep_data(data[train_idx])
    test_x, test_y = generate_multistep_data(data[test_idx])
    print(f"[INFO] Training Inputs Shape :: {train_x.shape}")
    print(f"[INFO] Training Outputs Shape :: {train_y.shape}")
    print(f"[INFO] Testing Inputs Shape :: {test_x.shape}")
    print(f"[INFO] Testing Outputs Shape :: {test_y.shape}")
    model.fit(
        train_x,
        train_y,
        validation_data=(test_x, test_y),
        epochs=25,
        batch_size=32,
        verbose=False,
        callbacks=callbacks,
    )
    history[i+1] = callbacks[1].history
    print()

[INFO] Working on Fold 1

---------------------------------------------------------------------------
KeyboardInterrupt                         Traceback (most recent call last)
Cell In[142], line 17
      7 os.makedirs(model_dir, exist_ok=True)
      8 callbacks = [
      9     ModelCheckpoint(
     10         f"{model_dir}/model.h5",
   (...)     15     TrainingCallback(train_idx, test_idx),
     16 ]
---> 17 train_x, train_y = generate_multistep_data(data[train_idx])
     18 test_x, test_y = generate_multistep_data(data[test_idx])
     19 print(f"[INFO] Training Inputs Shape :: {train_x.shape}")

Cell In[136], line 4, in generate_multistep_data(data)
      2 X_list, Y_list = [], []
      3 for p in range(data.shape[0]):
----> 4     X_combined, y_target = generate_patient_rolling_data(data[p])
      5     X_list.extend(X_combined)
      6     Y_list.extend(y_target)

Cell In[135], line 7, in generate_patient_rolling_data(data, static_feature_indices, input_steps, output_offset)
      5 total_steps = data.shape[0]
      6 all_indices = np.arange(num_features)
----> 7 temporal_indices = np.setdiff1d(all_indices, static_feature_indices)
      8 max_start = total_steps - input_steps - output_offset + 1
      9 X_list, Y_list = [], []

File f:\TfGpuEnv\Lib\site-packages\numpy\lib\arraysetops.py:979, in setdiff1d(ar1, ar2, assume_unique)
    977     ar1 = np.asarray(ar1).ravel()
    978 else:
--> 979     ar1 = unique(ar1)
    980     ar2 = unique(ar2)
    981 return ar1[in1d(ar1, ar2, assume_unique=True, invert=True)]

File f:\TfGpuEnv\Lib\site-packages\numpy\lib\arraysetops.py:274, in unique(ar, return_index, return_inverse, return_counts, axis, equal_nan)
    272 ar = np.asanyarray(ar)
    273 if axis is None:
--> 274     ret = _unique1d(ar, return_index, return_inverse, return_counts, 
    275                     equal_nan=equal_nan)
    276     return _unpack_tuple(ret)
    278 # axis was specified and not None

File f:\TfGpuEnv\Lib\site-packages\numpy\lib\arraysetops.py:336, in _unique1d(ar, return_index, return_inverse, return_counts, equal_nan)
    334     aux = ar[perm]
    335 else:
--> 336     ar.sort()
    337     aux = ar
    338 mask = np.empty(aux.shape, dtype=np.bool_)

KeyboardInterrupt:
In [ ]:
with open('models/callback.json', 'w') as f:
    json.dump(history, f, indent=4)
In [ ]:
def plot_loss(df, split):
    df_melt = df.melt(id_vars=["fold", "epochs"])
    g = sbn.FacetGrid(
        df_melt,
        col="variable",
        col_wrap=2,
        sharex=False,
        sharey=False,
        height=3,
        aspect=1.75,
    )
    g.map_dataframe(sbn.lineplot, x="epochs", y="value", hue="fold", palette="Set2")
    g.set_titles(col_template="horizon = {col_name}")
    for ax in g.axes.flat:
        ax.legend(ncol=5)
        ax.get_legend().set_title("Fold")
        ax.get_legend()._legend_box.align = "left"
        ax.set_xlabel('Epochs')
        ax.set_ylabel('MSE')
    g.fig.suptitle(f'MSE For {split.title()}ing Data In Cross Validation')
    g.fig.tight_layout()
    g.fig.savefig(f'models/{split}.png')
In [ ]:
with open('models/callback.json', 'r') as f:
    history = json.load(f)
    
for split in ['train', 'test']:
    dfs = []
    for fold in history:
        df = pd.DataFrame(history[fold][split])
        df.insert(0, 'epochs', range(1, len(df)  + 1))
        df.insert(0, 'fold', fold)
        dfs.append(df)
    df = pd.concat(dfs)
    df.to_csv
    plot_loss(df, split)
No description has been provided for this image
No description has been provided for this image
In [ ]:
def evaluate(y_true, y_pred):
    return [
        mean_squared_error(y_true, y_pred),
        root_mean_squared_error(y_true, y_pred),
        mean_absolute_error(y_true, y_pred),
        r2_score(y_true, y_pred),
    ]
In [ ]:
def evaluate_metrics_by_horizon(data, model, horizons=[3, 6, 9, 12]):
    y_preds = autoregressive_forecast(data, model)
    y_true = np.stack([data[:, 3 + h, :18] for h in range(4)], axis=1)
    metrics_per_horizon = {}
    for i, h in enumerate(horizons):
        metrics_per_horizon[f"{h}m"] = evaluate(y_true[:, i], y_preds[:, i])
    return metrics_per_horizon
In [ ]:
metrics = {'train': {}, 'test': {}}

for i, (train_idx, test_idx) in enumerate(kfold.split(np.zeros((data.shape[0], 1)))):
    print(f"[INFO] Evaluating on Fold {i+1}")
    model = build_transformer_model()
    model.load_weights(f'models/fold_{i+1}/model.h5')
    callbacks = [
        ModelCheckpoint(
            f"{model_dir}/model.h5",
            save_best_only=True,
            save_weights_only=True,
            verbose=False,
        ),
        TrainingCallback(train_idx, test_idx),
    ]
    metrics['train'][i+1] = evaluate_metrics_by_horizon(data[train_idx], model)
    metrics['test'][i+1] = evaluate_metrics_by_horizon(data[test_idx], model)

[INFO] Evaluating on Fold 1
[INFO] Evaluating on Fold 2
[INFO] Evaluating on Fold 3
[INFO] Evaluating on Fold 4
[INFO] Evaluating on Fold 5
In [ ]:
os.makedirs('results', exist_ok=True)
In [ ]:
METRICS = ['MSE', 'RMSE', "MAE", 'R2-Score']

metrics_df = []
for split in ['train', 'test']:
    dfs = []
    for fold in metrics[split]:
        df = pd.DataFrame(metrics[split][fold])
        df.insert(0, 'Metrics', METRICS)
        df.insert(0, 'Fold', fold)
        dfs.append(df)
    df = pd.concat(dfs)
    df.insert(0, 'Split', split.title())
    metrics_df.append(df)

metrics_df = pd.concat(metrics_df)
metrics_df.to_csv('results/all_metrics.csv', index=False)
In [ ]:
def plot_metrics(df, metric):
    ylims = {'MSE': [1.5, 5], 'RMSE': [0.6, 1.5], 'MAE': [0.75, 1], 'R2-Score': [0.3, 1]}
    df = df[df['Metrics'] == metric].copy()
    df.drop('Metrics', axis=1, inplace=True)
    df_melt = df.melt(id_vars=['Fold', 'Split'])
    g = sbn.FacetGrid(df_melt, col='Split', sharex=False, sharey=False, height=4, aspect=1.25)
    g.map_dataframe(sbn.barplot, x='variable', y='value', hue='Fold', palette='tab10')
    for ax in g.axes.flat:
        ax.grid(True)
        for p in ax.patches:
            p.set_edgecolor('k')
            p.set_linewidth(1)
            p.set_zorder(1000)
        ax.legend(ncol=3)
        ax.get_legend().set_title("Fold")
        ax.get_legend()._legend_box.align = "left"
        ax.set_xlabel('Horizon')
        ax.set_ylim(ylims[metric])
    g.fig.suptitle(f'Metric = {metric}')
    g.fig.tight_layout()
    g.fig.savefig(f'results/{metric}.png')
In [ ]:
for metric in METRICS:
    plot_metrics(metrics_df, metric)
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
In [ ]:
df_r2_test = metrics_df[
    (metrics_df["Metrics"] == "R2-Score") &
    (metrics_df["Split"] == "Test")
].copy()
df_r2_test["avg_r2"] = df_r2_test[["3m", "6m", "9m", "12m"]].mean(axis=1)
df_r2_test.reset_index(drop=True, inplace=True)
best_row = df_r2_test["avg_r2"].idxmax()
best_fold = df_r2_test.loc[best_row]['Fold']
print(f'Best Fold :: {best_fold}')

Best Fold :: 4
In [ ]:
fold_df = metrics_df[metrics_df['Fold'] == best_fold]
fold_df.drop('Fold', axis=1, inplace=True)
fold_df = fold_df.round(4)
fold_df.to_csv('results/best_metrics.csv', index=False)
fold_df_melt = fold_df.melt(id_vars=['Split', 'Metrics'])
g = sbn.FacetGrid(fold_df_melt, col='Split', sharex=False, sharey=False, height=4, aspect=1.25)
g.map_dataframe(sbn.barplot, x='variable', y='value', hue='Metrics', palette='husl')
for ax in g.axes.flat:
    ax.grid(True)
    for p in ax.patches:
        p.set_edgecolor('k')
        p.set_linewidth(1)
        p.set_zorder(1000)
    ax.legend(ncol=2)
    ax.get_legend().set_title("Metrics")
    ax.get_legend()._legend_box.align = "left"
    ax.set_xlabel('Horizon')
    ax.set_ylim(0, 4)
g.fig.suptitle('Best Fold Metrics Across All Horizons')
g.fig.tight_layout()
g.fig.savefig('results/best_metrics.png')
No description has been provided for this image
In [ ]:
def autoregressive_forecast_multiple_horizons(
    model,
    patient_data,
    static_feature_indices=[18, 19, 20, 21],
    input_steps=4,
    horizons=[3, 6, 9, 12],
):
    temporal_indices = np.setdiff1d(
        np.arange(patient_data.shape[1]), static_feature_indices
    )
    x_seq = patient_data[:input_steps].copy()
    static_values = patient_data[0, static_feature_indices]
    pred_outputs = []
    all_preds = []
    for month in range(1, max(horizons) + 1):
        static_repeated = np.repeat(static_values[np.newaxis, :], input_steps, axis=0)
        x_model = np.concatenate([x_seq[:, temporal_indices], static_repeated], axis=1)
        x_model = np.expand_dims(x_model, axis=0)
        pred = model.predict(x_model, verbose=0)[0]
        all_preds.append(pred)
        x_next = np.concatenate([pred, static_values])
        x_seq = np.concatenate([x_seq[1:], [x_next]], axis=0)
        if month in horizons:
            pred_outputs.append(pred)
    return np.array(pred_outputs)
In [143]:
def plot_actual_vs_pred_for_patient(
    pid, static_feature_indices=[18, 19, 20, 21], input_steps=4, horizons=[3, 6, 9, 12]
):
    for i, (_, test_idx) in enumerate(kfold.split(np.zeros((data.shape[0], 1)))):
        if best_fold == i + 1:
            patient_data = data[test_idx][pid]
            y_pred = autoregressive_forecast_multiple_horizons(model, patient_data)[
                :, 10
            ].ravel()
            y_pred = np.array([*patient_data[:4, 10].ravel(), *y_pred])
            temporal_indices = np.setdiff1d(
                np.arange(patient_data.shape[1]), static_feature_indices
            )
            y_actual = []
            for h in horizons:
                idx = input_steps + h - 1
                y_actual.append(patient_data[idx, temporal_indices])
            y_actual = np.array(y_actual)[:, 10]
            y_actual = np.array([*patient_data[:4, 10].ravel(), *y_actual])
            x = [0, 1, 2, 3, 6, 9, 12, 15]
            fig = plt.figure(figsize=(8, 4))
            ax = fig.gca()
            ax.plot(x, y_actual, c="green", marker="o", label="Actual")
            ax.plot(x, y_pred, ls='--', c="blue", marker="s", label="Predicted")
            ax.plot(x[:4], y_actual[:4], c="k", marker="o", label="First 3 Months")
            ax.set_title(f"Patient #{pid + 1}")
            ax.set_ylabel("ALSFRS Total Score")
            ax.grid(True)
            ax.set_xticks(
                [0, 1, 2, 3, 6, 9, 12, 15],
                [
                    "0\n(1st visit)",
                    "1",
                    "2",
                    "3",
                    "6\n(+3 months)",
                    "9\n(+6 months)",
                    "12\n(+9 months)",
                    "15\n(+12 months)",
                ],
            )
            ax.legend()
            fig.tight_layout()
            fig.savefig(f'results/predicted/{pid + 1}.png')
In [146]:
import tqdm

os.makedirs("results/predicted", exist_ok=True)
for pid in tqdm.tqdm(range(100), desc="[INFO] Forecasting Patient ALS Progression :"):
    try:
        plot_actual_vs_pred_for_patient(pid)
    except Exception:
        break

[INFO] Forecasting Patient ALS Progression :: 100%|██████████| 100/100 [02:57<00:00,  1.78s/it]
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image
No description has been provided for this image